ChatGPT Ads ROAS Study: What $10,000 of Real Spend Actually Returns

Henry Purchase

Co-Founder

Every ChatGPT Ads ROAS article online right now is theory.

They walk you through how to calculate ROAS. They give you hypothetical numbers ("imagine you spend $5,000 and make $20,000"). They explain attribution models. What none of them give you is a single real figure from money actually spent on the platform, because almost nobody is sharing real data.

So I am running the study myself. I am spending $10,000 of my own money on ChatGPT Ads across multiple campaigns, tracking every dollar in and every dollar out, and publishing the results here as they land.

This is a living document. The methodology and early signals are below now. The hard ROAS numbers fill in as the data matures over the coming weeks. Bookmark it and check back, or join the community where I post the weekly updates first.

What this study measures

The goal is a single honest question: does ChatGPT Ads return more than it costs, and for whom?

To answer it properly, I am tracking:

• Spend — total and per campaign
• Clicks and CPC — what I am actually paying per click
• Conversions — leads and bookings, tracked through the OpenAI pixel and Conversions API
• Cost per conversion — the number that decides whether a campaign lives or dies
• Direct ROAS — revenue directly attributed to ChatGPT ad clicks
• Blended ROAS — direct plus assisted conversions, since a lot of ChatGPT-influenced buyers convert later through other channels

That last distinction matters more on ChatGPT than anywhere else, and I will explain why below.

Why ChatGPT Ads ROAS is hard to measure (and why most numbers will be wrong)

Before any results, you need to understand why a naive ROAS number on this platform is misleading.

ChatGPT is a research environment. People do not click an ad and buy thirty seconds later the way they sometimes do on Google Shopping. They click, they read, they go back to ChatGPT and ask more questions, they leave, and they come back days later through a branded search or a direct visit. The conversation is the start of a research journey, not the end of a buying decision.

This creates a systematic measurement problem. Last-click ROAS will understate ChatGPT Ads' true return, often by half. If you only count same-session conversions, you will conclude the channel does not work and cut it, when it may be quietly driving conversions that show up later under "direct" or "organic" in your analytics.

That is why this study tracks two numbers:

Direct ROAS counts only conversions attributed directly to a ChatGPT ad click within the attribution window. It is conservative and clean.

Blended ROAS adds the assisted conversions: the branded search lift, the direct traffic bump, and the multi-touch journeys that started with a ChatGPT ad. It is closer to the truth but requires more careful attribution.

A real example of the gap, using round numbers: a campaign that looks like 2.1x ROAS on last-click can be 4.2x once you account for assisted conversions. That is the difference between cutting a channel and scaling it. Getting this right is the entire point of the study.

If your tracking is not set up to capture this properly, the numbers in this study will not be reproducible for you. Fix that first with the ChatGPT Ads tracking guide.

The methodology

Here is exactly how the study is structured so you can judge the data fairly and replicate it.

The reason for splitting ad groups by targeting method is that the study is not just measuring "does ChatGPT Ads work" but "which targeting approach works best." Broad context hints versus tight, buying-stage hints versus comparison-stage hints. Different ad groups isolate each so the winning approach is visible in the data, not guessed.

Early signals (updated as data lands)

This section updates weekly. Here is where things stand right now.

On cost. Against $3 to $5 max bids, CPCs are tracking inside that range so far, in line with the wider market. For context, CPMs across the platform have dropped from $60 at launch in February to as low as $25 now, so the channel is getting cheaper as more inventory opens up. Early movers are buying clicks before the auction gets crowded. I will publish my own observed CPC range here once I have enough volume for it to be representative rather than anecdotal.

On the auction. The relevance-weighted, second-price auction is real and it rewards tight targeting. Early on, ad groups with specific context hints and on-message creative are winning impressions at lower effective CPCs than broad ad groups with higher bids. The exact size of that relevance effect goes here once I have enough matched comparisons to quantify it fairly.

On conversions. Cost per conversion by ad group is still accumulating. I will not post a number until each ad group has enough conversions behind it to mean something. Those figures land in this section as they firm up.

On the targeting test. The study runs broad, high-intent, and comparison ad groups side by side specifically to see which targeting method produces the lowest cost per conversion. The winner gets called here once the data separates clearly.

On ROAS. The headline numbers, direct ROAS and blended ROAS, are the last thing I will publish, because they need the full 30-day window and enough conversions to be real. They go here when they are solid.

I am not going to publish a ROAS figure until there is enough conversion data to make it meaningful. A ROAS number off three conversions is noise, and this study is meant to be the opposite of the speculative numbers already out there. When the figures are solid, they go here.

What I expected going in (and where the surprises are landing)

A few hypotheses I started the study with. I will mark each one confirmed, rejected, or mixed as the evidence comes in, including the ones I get wrong.

Hypothesis 1: ChatGPT Ads convert better than cold Google traffic because intent is richer. The theory is that someone describing their full problem to ChatGPT is further down the decision path than someone typing three words into Google. Status: testing. The verdict and the evidence behind it land here.

Hypothesis 2: The Free and Go tier audience converts on value-first offers, not hard sells. Ads only show to Free and Go tier users, who skew toward lower immediate buying intent than the Plus and Pro users who see no ads. My bet was that lead magnets and free offers would outperform direct-purchase asks. Status: testing. Result here as the offer split resolves.

Hypothesis 3: Tight context hints beat broad ones on cost per conversion. This is the core of the targeting test above. Status: testing. Result here once the ad groups separate.

Hypothesis 4: A meaningful share of conversions arrive late, not same-session. If true, this is the most important finding in the study, because it means most people measuring last-click ROAS are undercounting and will wrongly conclude the channel fails. Status: testing, and the one I am watching most closely. The assisted-conversion share goes here.

Updating these honestly, including the ones I get wrong, is the whole value of a real study versus a theoretical one.

What counts as a good ChatGPT Ads ROAS?

People searching for this study want a benchmark. Here is the honest answer.

There is no established ChatGPT Ads ROAS benchmark yet. The platform is barely three months into self-serve, OpenAI has not published average ROAS, CPA, or CTR figures by vertical, and anyone quoting a confident category benchmark is guessing. Treat any "expected ROAS" number you see, including provisional ones in this study, as directional, not gospel.

What you can do is set your own threshold based on your other channels. If your Google Ads run at 4x ROAS, that is your bar. If ChatGPT Ads beat it, scale. If they come in lower but the assisted-conversion data shows a strong late-conversion tail, give it more time before judging. If they come in well below with no assisted lift, the channel may not fit your business yet, and that is a legitimate finding too.

The point of this study is to give you a real reference point from real spend, so you are setting expectations against data instead of hype.

How to run your own ChatGPT Ads ROAS test

If you want to run your own version of this study rather than wait for mine, here is the short framework:

1. Set a learning budget, not a performance budget. $500 to $2,000 over two to four weeks. Enough to generate meaningful data, small enough that a loss is tuition, not damage.
2. Wire up tracking before you spend a dollar. Pixel plus Conversions API plus UTMs. Without this, your ROAS number is fiction. Here is the full setup.
3. Split ad groups by targeting method so you can see which approach wins, not just whether the channel works overall.
4. Measure two ROAS numbers, direct and blended, so you do not undercount late conversions.
5. Give it 14 days minimum before any judgement and 30 days before a scale-or-kill decision. ChatGPT Ads need time to stabilise.
6. Compare against your existing channels' ROAS, not against an imaginary benchmark.

Reading the data is the hard part, because the ChatGPT dashboard makes cross-campaign analysis painful. The workaround I use is exporting the CSV and analysing it in Claude. That manual process is what we are building Focal to replace: connect your ChatGPT Ads account and ask "what is my real ROAS by ad group and where am I wasting spend" from inside Claude or ChatGPT, and get the answer without the spreadsheet gymnastics.

Follow the study

This page updates as the data lands. The fastest way to follow along:

• FutureProof community — where I post the weekly numbers and the full breakdowns first, with 350+ operators comparing their own results
• YouTube channel — video breakdowns of the study as it progresses, including the campaigns and creative behind the numbers
• Focal waitlist — for the tool that turns this kind of analysis from a weekly manual job into a question you ask in plain language

I will keep spending and keep reporting, including the parts that do not go to plan. That is the only kind of ROAS study worth reading.

Frequently asked questions

What is a good ROAS for ChatGPT Ads?

There is no established benchmark yet. The platform is too new and OpenAI has not published ROAS figures by vertical. Set your threshold against your existing channels: if your Google Ads run at 4x, use that as your bar. Treat any quoted "expected ChatGPT Ads ROAS" as directional only.

Do ChatGPT Ads actually deliver positive ROAS?

That is exactly what this study is measuring with real spend. The honest answer today is that it depends heavily on your category, your offer, and crucially how you measure. Last-click ROAS understates the channel because many ChatGPT-influenced conversions arrive later through other channels. Blended ROAS that captures assisted conversions tells a truer story.

Why is my ChatGPT Ads ROAS lower than my Google Ads ROAS?

Two likely reasons. First, you may be measuring last-click only and undercounting late conversions, which hits ChatGPT harder than Google because ChatGPT is a research environment. Second, the audience is Free and Go tier users with different intent than Google searchers, so offers that work on Google may need reframing as value-first on ChatGPT.

How much should I spend to test ChatGPT Ads ROAS?

$500 to $2,000 over two to four weeks for a personal test. Treat it as a learning budget, not a performance budget. You need enough volume to generate meaningful data but not so much that a loss hurts.

How long until I can trust my ChatGPT Ads ROAS number?

At least 14 days for a directional read, 30 days for a decision. ChatGPT Ads need two to four weeks to stabilise while the platform learns your audience, and the late-conversion tail means early numbers undercount.

How do I track ROAS on ChatGPT Ads accurately?

Use the OpenAI pixel and Conversions API together, add UTM parameters to every ad URL, and measure both direct and blended ROAS. The full tracking setup is here.

My question is not listed here.

Ask in the FutureProof community or email support@tryfocal.com.